# Microsoft’s New Phi 4 AI Family Packs a Punch, Rivals Larger Models

## Microsoft’s New Phi 4 AI Family Packs a Punch, Rivals Larger Models

Microsoft has unveiled its latest generation of “open” AI models, the Phi 4 family, boasting impressive performance relative to their size. The new models, announced Wednesday, aim to provide developers with a foundation for building AI applications, particularly for resource-constrained environments.

The Phi 4 lineup includes three permissively licensed models: Phi 4 mini reasoning, Phi 4 reasoning, and Phi 4 reasoning plus. These models are specifically designed for “reasoning,” enabling them to dedicate more computational resources to fact-checking and problem-solving. This focus allows them to tackle complex tasks more effectively.

The smallest of the trio, Phi 4 mini reasoning, is a 3.8-billion-parameter model trained on synthetic math problems generated by DeepSeek’s R1 reasoning model. Microsoft envisions this model primarily for educational purposes, highlighting its potential for “embedded tutoring” on lightweight devices.

Phi 4 reasoning steps up to 14 billion parameters and is trained on a combination of high-quality web data and curated examples from OpenAI’s o3-mini. This model excels in math, science, and coding-related applications, according to Microsoft.

Finally, Phi 4 reasoning plus is an adaptation of Microsoft’s existing Phi-4 model, refined for enhanced reasoning capabilities and accuracy. Impressively, Microsoft claims that Phi 4 reasoning plus approaches the performance levels of DeepSeek’s R1, despite R1 having a significantly larger parameter count of 671 billion. Internal benchmarks also show Phi 4 reasoning plus matching OpenAI’s o3-mini on the OmniMath benchmark.

The launch of the Phi 4 family underscores Microsoft’s commitment to developing efficient and accessible AI models. As Microsoft stated in a blog post, these models “balance size and performance” through the use of techniques like distillation, reinforcement learning, and high-quality data. This approach allows even devices with limited resources to handle complex reasoning tasks efficiently.

All three models – Phi 4 mini reasoning, Phi 4 reasoning, and Phi 4 reasoning plus – are available on the AI development platform Hugging Face, accompanied by detailed technical documentation. Developers can now explore and leverage these powerful, yet compact, AI models for a wide range of applications.